The three blocks under the floor: who runs what (F2, where the Claude Code lens lives in full), when the floor is busy (F3) and what leaves the building (F4, the loading dock). Each is drawn two ways, on the same world and controls as the eight directions. Click any cell, row or band to focus it.
Lens
Harness
Period
Scale
State
Who is working, and where? · lit = worked with that harness in the period · dim = installed, idle · dashed amber = detected only, no sensor · red = errored · the harness picker always drives F2b
200 people · 248 stations · 6 harnesses — 19 stations unseen, Hermes on 9, 2 sensitive exits
F2
Who runs what
Teams × harnesses: where Claude Code, Cursor, Codex… are actually used, and by whom.
Form F2a
Team × harness matrix
Strong at precision — every cell a number, every lens a column you can read across; the Hermes column says "detected" instead of pretending to a token count.Weak at image — a table with tone; at 41 teams × 7 harnesses it is a spreadsheet you scan, not a picture you see.
Team
Claude Code
Cursor
Codex
Copilot
Gemini CLI
Windsurf
Hermes Agent
Team
Platform 28
22
412.6M
4
136.8M
6
50.3M
2
3.9M
2
7.7M
4
17.7M
·
628.8M
Payments 23
5
14.3M
15
218.7M
1
8.2M
1
879k
2
9.3M
·
1
detected
251.4M
ML Research 17
10
301.1M
5
148.5M
4
16.1M
2
24.4M
1
6.3M
·
3
detected
496.4M
Mobile 19
6
12.4M
14
124.1M
1
8.4M
2
18.7M
·
1
292k
·
163.9M
Data Eng 16
3
18.7M
2
14.7M
6
92.1M
1
3.8M
1
6.0M
·
2
detected
135.3M
Web 21
16
270.7M
4
11.1M
4
45.0M
3
3.3M
1
10.9M
1
7.3M
·
348.2M
Support 17
2
23.3M
1
224k
0
0
6
78.2M
·
0
0
·
101.7M
Security 8
4
21.5M
·
1
5.8M
2
7.5M
·
1
399k
·
35.1M
Sales Eng 14
1
3.5M
1
897k
1
319k
3
49.6M
·
·
3
detected
54.3M
Finance 8
1
9.2M
·
1
5.3M
2
22.5M
·
·
·
37.0M
Legal 6
·
·
·
·
·
·
·
0
Design 10
1
16.1M
·
1
26.7M
·
·
4
107.2M
·
149.9M
Unassigned 13
7
142.0M
4
15.2M
·
2
47.0M
·
·
·
204.2M
cell = people who worked with the harness in the period (tokens below) · tone = share of the team's peopleclick a column header to pick that harness for F2bdetected only (Hermes)
Form F2b
The harness lens — Claude Code
Strong at one harness, fully: adoption per team on one headcount scale, idle and detected-only people kept visible, versions, share of tokens, findings. Mark's Claude Code question answered in one block.Weak at the other harnesses — it is a drill-in, one harness at a time.
Claude Code
102 installs · 95 people · 12 teams · 48% of the floor's tokens · 7 detected only · 1 errored
Versions across 102 installs
2.5.2
2.4.0
2.3.1
TeamPeople — using it · installed, idle · detected only · grey = not using (bar length = headcount)Of team tokensFindings · erroredPlatform66%2 open · 1 errPayments6%1 openML Research61%1 openMobile8%—Data Eng14%—Web78%—Support23%—Security61%—Sales Eng6%—Finance25%—Legal——Design11%—Unassigned70%—
F3
Shifts — when the floor is busy
Team × hour of day (org-local), summed across 30d.
Form F3a
Heat strip (one scale for the floor)
Strong at when the floor is busy — one scale, so big teams at 2pm glow and small teams stay pale; the Floor row on top is the company's day.Weak at small teams' shape — a quiet team is a pale row even if it works nights.
tone = tokens in that hour, one scale for every team (the Floor row on its own scale) · white-to-grape, square root
Form F3b
Punchcard (each team's own day)
Strong at each team's own day — shape only, so a small team's late shift is as visible as Platform's; the punchcard grammar reads without a legend.Weak at how much — every row is scaled to itself, so it can't say which team is busy.
dot area = share of that team's own day (each row on its own scale) — shape, not size; the heat strip carries size
F4
The loading dock — what leaves the building
The brand's first question at floor scale: where is my data going? The floor on the left, the doors on the right — AI providers on company accounts, personal accounts as their own door, and the web the agents reach. Width = volume; each band is stacked by data class, sensitive classes in amber (customer data, PII) and rose (secrets).
Form F4a
Dock doors
Strong at the picture — the floor on the left, doors on the right, bands you can follow; Personal accounts and pastebin read as doors that shouldn't be there, even at floor width.Weak at small doors: at linear width a sensitive trickle is a hairline; the outline and the event count carry it.
Source codeDocs & ticketsCustomer dataPIISecrets & keyssensitive events through a door the company doesn't holdfloor width, or reached only from detected-only machinesvolume = token-equivalent; web destinations are estimated from agent reach
Form F4b
Dock manifest
Strong at reading the numbers — one linear scale, plus a second bar for the sensitive share of each door's own volume, so a 6M personal-account door that is 20% sensitive outranks a 1.7B provider that is 8%.Weak at the "leaving the building" feeling — it is a ledger, not a place.
DoorVolume by class · one linear scaleVolumeSensitive share of the doorAI providers · company accountsAnthropic1.7B8.4%11 eventsOpenAI513.8M8.4%10 eventsAzure OpenAI193.7M8.4%3 eventsGoogle135.1M8.4%8 eventsMistral59.7M8.4%OpenRouter9.6M8.4%Personal accounts · accounts the company doesn't holdPersonal accounts5.9M20%2 eventsWeb destinations · what the agents reachgithub.com209.7M0%registry.npmjs.org78.7M0%pypi.org52.4M0%slack.com39.3M20%pastebin.com5.2M50%3 events
Source codeDocs & ticketsCustomer dataPIISecrets & keyssensitive events through a door the company doesn't holdfloor width, or reached only from detected-only machinesvolume = token-equivalent; web destinations are estimated from agent reach
Read
My read
F2: ship both, stacked
The matrix is the overview, the harness lens is the drill-in, and they share one control (click a column header). The harness lens is where Mark's Claude Code question lands: adoption per team, who installed it and stopped, which machines we can't see, and how far behind latest the fleet is.
F3: the heat strip, with the punchcard as its toggle
The page's question is "when is the floor busy", which needs one scale. The punchcard answers "when does each team work" — worth one click, not a second block. Neither can carry violations or coverage until an hourly series exists.
F4: the doors picture on /fleet, the manifest in the drawer
Dock doors is the image the brief asks for and the one a security engineer screenshots. The manifest's sensitive-share bar fixes the doors' one weakness, so it belongs one click away (the focused door already shows its class mix).
The dock is whole on day one
Every other block needs teams mapped before it says much; the dock doesn't. That makes it the first-run hero for /fleet — "here is what already leaves the building" — while the floor waits for teams.
Judge
What to try, and what to look for
Pick Harness → Claude Code, then Codex
F2b should re-draw completely: adoption bars, versions, share. Does it answer "how far has it spread, and where is it stuck" without the matrix?
Lens → Violations
The matrix turns to status ink; the dock mutes ordinary data so only customer data, PII and secrets stay lit. Does the dock now read as the security view?
State → All clear
The Personal accounts door and pastebin.com disappear, sensitive classes drop to zero at the AI doors, Hermes leaves the matrix. The dock should look calm, not empty.
Scale → 1,500
The matrix becomes 41 rows, the heat strip compresses rows to 13px, the dock doesn't change shape at all (doors don't scale with people). Which block still reads first?
Period → 24h
Shifts become today's shape. The dock thins; do the sensitive doors still show? They should — floor width plus the outline.
Click a dock band
The door focuses and its class mix opens underneath. On the real page this is the provider detail (company accounts), the personal-account finding, or the egress host.